Edge Video Quality Monitoring Using Hierarchical ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video quality monitoring systems face logistical and resource challenges in assessing delivered video content against reference content, making it difficult to replicate human perception of image quality using machine analysis across large-scale video networks.
Innovation Solution
A video quality monitoring system that leverages machine learning models integrated within edge devices like set-top boxes and smart televisions to analyze video content across multiple domains, including IP, MPEG transports, video and audio bitstreams, pixels, and metadata, allowing for real-time quality assessment without relying on comparisons to reference images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If video quality is assessed by comparing reference pictures against delivered video content, then measurement precision is improved, but device complexity and resource requirements worsen
Solution Approach 1:
The patent extracts the reference picture comparison function from the client device and relocates it to the headend server. The headend generates synthetic reference pictures using machine learning models and sends only the quality metrics to clients, eliminating the need for clients to store and compare large volumes of reference video data locally.
Solution Approach 2:
The patent introduces a headend server as an intermediary between the video source and client devices. This headend performs the complex reference picture generation and comparison operations centrally, acting as a mediator that simplifies the client device architecture while maintaining accurate quality assessment through server-side processing.
2Measurement precision
If reference video content is transmitted to user locations for comparison, then measurement precision is improved, but network bandwidth usage and loss of substance worsen
Solution Approach 1:
The patent extracts only the essential quality assessment data from the reference comparison process and transmits it to clients, rather than transmitting entire reference video streams. This extraction approach maintains measurement precision while dramatically reducing network bandwidth consumption.
Solution Approach 2:
The patent creates synthetic reference pictures using machine learning models at the headend, replacing the need to transmit actual reference video content. These synthesized references are generated on-demand and used for quality comparison, eliminating bandwidth-intensive reference stream transmission.
3Productivity
If machine learning models analyze video content at edge devices, then productivity is improved, but device complexity and use of energy worsen
Solution Approach 1:
The patent implements a hybrid approach where edge devices perform partial machine learning analysis on local video content, while the headend performs more intensive processing on synthesized reference comparisons. This partial action at the edge provides real-time responsiveness without the full energy cost of complete local analysis.
Solution Approach 2:
The system implements feedback loops where quality metrics from edge device analysis are sent to the headend, which then adjusts and refines the machine learning models. This feedback mechanism allows edge devices to use pre-trained models for efficient local analysis, reducing their energy consumption while maintaining high productivity through centralized model optimization.
Data Source
AI summary
A device is provided that includes computer-readable storage media storing one or more sequences of instructions and processing circuitry configured to execute the one or more sequences of instructions. Upon executing the instructions, the processing circuitry may receive network packets containing content encapsulated in multiple layers; process the received network packets to extract the content for presentation; generate a predicted presentation quality indicator for the extracted content using machine learning models in a hierarchical order with data generated during processing of the received network packets used as inputs to the machine learning models; and provide the predicted presentation quality indicator for the extracted content to a server via a network, wherein the data generated during processing of the received network packets is correlated across the layers to generate the predicted presentation quality indicator.